bio-clip-seq-m6a-clip

What this skill does

## Version Compatibility

Reference examples tested with: miCLIP2 pipeline (Kortel 2021), m6Aboost 1.0+, GLORI-tools (Liu 2023), Bullseye 1.0+, m6Anet 2.1+, EpiNano 1.2+, MeRIPSeq tools (exomePeak2 1.16+), nanocompore 1.0+, samtools 1.19+, bedtools 2.31+, R 4.3+.

Before using code patterns, verify installed versions match. If versions differ:
- Python: `pip show <package>` then `help(module.function)` to check signatures
- R: `packageVersion('<pkg>')` then `?function_name` to verify parameters
- CLI: `<tool> --version` then `<tool> --help` to confirm flags

If code throws unexpected errors, introspect the installed package and adapt the example to match the actual API rather than retrying.

# m6A CLIP (N6-Methyladenosine Profiling)

**"Map m6A modifications at single-nucleotide resolution"** -> Profile m6A on RNA using one of three orthogonal approaches: antibody-based UV-CL (miCLIP/miCLIP2), antibody-free chemical conversion (GLORI), or enzyme-fusion editing (DART-seq with APOBEC1-YTH). Nanopore direct RNA (m6Anet, nanocompore, EpiNano) provides a fourth modality. The DRACH consensus motif (D=A/G/U, R=A/G, A=m6A, C=C, H=A/C/U) constrains plausible sites but is not exclusive - only a fraction of DRACH instances are methylated; some m6A sites occur outside DRACH. Cross-method discordance is real: DART-seq C->U mutations are 44% within DRACH motifs (Liu 2023), suggesting many DART sites are not consensus m6A. GLORI is the new (2023) gold standard for stoichiometric single-base m6A.

- CLI (miCLIP2 antibody-based): `iCount` or custom pipeline through truncation + C->T mutation analysis; then m6Aboost ML scoring
- CLI (GLORI antibody-free): `GLORI-tools` Python pipeline; output is per-A m6A fraction (stoichiometric)
- CLI (DART-seq editing): `Bullseye` or `SAILOR` pipeline; identify C->U editing sites; filter by DRACH; cross-check against APOBEC1-only control
- CLI (m6Anet nanopore): `m6anet inference` on nanopolish eventalign output; per-site probability of m6A
- CLI (MeRIP-seq peak calling): `exomePeak2` in R for peak-level m6A from IP+input MeRIP libraries

The m6A field is rapidly evolving (2022-2026); single-base methods (GLORI, m6Anet) have largely replaced antibody-based miCLIP for new studies, but miCLIP2 remains the most common because of its eCLIP-like processing pipeline. Cross-method discordance means high-confidence m6A reporting should require concordance across at least two orthogonal methods.

## Methods Taxonomy

| Method | Detection chemistry | Resolution | Antibody | Stoichiometry | Strength | Fails when |
|--------|---------------------|------------|----------|---------------|----------|------------|
| MeRIP-seq (Dominissini 2012, Meyer 2012) | Anti-m6A IP + RNA-seq | Peak (50-300 nt) | Yes | No | Original m6A method; widely used | Low resolution; cannot distinguish m6A from m6Am |
| miCLIP (Linder 2015) | Anti-m6A + UV-CL + RT mutation | Single-nucleotide (some) | Yes | No | Single-nt subset of m6A peaks | Low yield of single-nt; high false-positive rate |
| miCLIP2 (Kortel 2021) | Anti-m6A + UV-CL + improved library | Single-nucleotide | Yes | No | Higher complexity; ML-classified (m6Aboost) | Antibody specificity remains issue |
| GLORI (Liu 2023) | Glyoxal + nitrite chemical conversion of A to N1-methyl-2-amino-6-oxopurine | Single-nucleotide | No (chemical) | Yes (stoichiometric) | Stoichiometric m6A fraction per site | New; less validated; harsh conversion may damage rare RNAs |
| DART-seq (Meyer 2019) | APOBEC1-YTH fusion edits C adjacent to m6A | Single-nucleotide (offset) | No | No | Antibody-free; in vivo | Only 44% of edits in DRACH motifs; high false positive |
| m6A-CLIP (Ke 2015) | Anti-m6A + UV-CL | Peak | Yes | No | Original UV-CL approach | Predecessor to miCLIP |
| m6Anet (Hendra 2022) | Nanopore direct RNA + neural net | Single-nucleotide (DRACH constraint) | No | Probability | Direct RNA; preserves isoform context | Restricted to DRACH; needs high coverage per site |
| EpiNano (Liu 2019) | Nanopore + SVM on signal features | Single-nucleotide | No | No | Pioneer nanopore m6A | Lower accuracy than m6Anet on benchmark |
| nanocompore (Leger 2021) | Nanopore + statistical test wt vs Mettl3-KO | Single-nucleotide | No | No | Comparative; high specificity | Requires KO control sample |
| DENA (Zhong 2024) | Nanopore + transformer model | Single-nucleotide | No | No | Single-sample tool | Newer; less validation |
| FTO/ALKBH5-aware methods | Eraser perturbation | Site | No | Indirect | Validates m6A regulation | Indirect |
| MAZTER-seq (Garcia-Campos 2019) | RNase T1 cleavage at unmodified ACA | Site (within ACA) | No | No | Antibody-free | Restricted to ACA context (subset of DRACH) |
| REF-seq (Werner 2020) | Endonuclease-cleavage | Site | No | No | Antibody-free | Restricted context |
| m6ACali (Yang 2024) | Calibrates MeRIP / miCLIP | Site | NA | Yes (calibration) | Cross-method calibration | Postprocessing only |

Methodology evolves; verify the latest benchmark publications (e.g., Tegowski 2024 Mol Cell review). The field is moving toward GLORI as the new gold standard but miCLIP2 remains the most-cited method because of its eCLIP-pipeline compatibility.

## Critical Choice: Antibody-Based vs Antibody-Free

**Antibody-based (MeRIP-seq, miCLIP, miCLIP2, m6A-CLIP):** Anti-m6A antibody (Abcam/Synaptic Systems) immunoprecipitates m6A-bearing RNA. The antibody is the only limitation - false positives from non-specific binding to long structured RNAs (especially poly-A) and false negatives at sites with low m6A stoichiometry. Mettl3 knockout calibration is recommended.

**Antibody-free chemical (GLORI):** Glyoxal + nitrite converts unmodified A to a nucleotide that reads as G; m6A is protected and reads as A. Sites are detected as A->G discrepancies post-conversion. Stoichiometric (the fraction of reads showing A vs G at a position = m6A fraction). Most rigorous but chemistry is harsh - degrades very long RNAs.

**Antibody-free enzymatic (DART-seq, APOBEC1-YTH):** APOBEC1 cytidine deaminase fused to YTH-domain (m6A reader) edits C residues adjacent to m6A. Editing pattern (C->U) marks m6A nearby but not exactly. 44% of DART edits in DRACH; many edits are off-target.

**Antibody-free nanopore (m6Anet, nanocompore, EpiNano):** Direct RNA sequencing detects m6A via current signal perturbation. Preserves isoform context. m6Anet AUC 0.83 on HEK293T; outperforms EpiNano and Tombo on benchmark (Liu 2024).

| Goal | Method |
|------|--------|
| Stoichiometric m6A fraction per site | GLORI |
| eCLIP-compatible processing pipeline | miCLIP2 + m6Aboost |
| Isoform-resolved m6A | m6Anet (nanopore) |
| Cell-line comparison (KO available) | nanocompore vs Mettl3-KO |
| High-throughput screening | DART-seq (in vivo, no UV) |
| Initial discovery (low cost) | MeRIP-seq (with calibration) |
| Variants in m6A context | GLORI + variant-effect analysis |
| Combined m6A + 5'-cap m6Am | miCLIP2 (detects both with separate motifs) |

## DRACH Motif Constraint

The DRACH consensus (D=A/G/U, R=A/G, A=m6A, C=C, H=A/C/U) is the dominant motif at m6A sites - 70-90% of high-confidence sites fall in DRACH context. But:
- Some m6A sites occur outside DRACH (~10-20% in calibrated datasets)
- Many DRACH instances are NOT methylated (only a subset)
- Filtering for DRACH-only loses 10-20% of sites; not-filtering inflates false positives

**miCLIP2 + m6Aboost (Kortel 2021)** trained on Mettl3 knockout calibration data to score sites without DRACH filtering. The m6Aboost ML model is the recommended approach when DRACH-blind detection is needed.

**GLORI** does not filter by DRACH; the per-A m6A fraction is reported regardless of context. The non-DRACH GLORI sites (10-20%) include genuine m6A in non-canonical context.

## Cross-Method Discordance

| Comparison | Concordance | Source |
|------------|-------------|--------|
| miCLIP vs miCLIP2 | ~70% | Kortel 2021 |
| miCLIP2 vs GLORI | ~60% (miCLIP2 calls in GLORI) | Liu 2023 |
| GLORI vs MeRIP-seq peaks |